Diabetes and CHD were clinically verified (Alberti and Zimmet, 1998 and Ferrie et al., 2006). In descriptive analyses, we evaluated variables across physical activity and mental health categories. Differences between the groups were tested by chi-square for categorical variables and ANOVA for continuous variables. Provisional analyses considering each outcome separately explored potential effects of cumulative exposure to one variable on the outcome of the other at end of follow-up using linear regression. Latent growth curve models allow participants with incomplete follow-up data

to be included in the analysis by acknowledging that repeated measures on the same individual are correlated (Bollen and Curran, 2006). The maximum likelihood ratio (MLR) estimator allows for moderate non-normality in continuous outcomes. The intercepts represent initial status at baseline (1997/99) for each variable. The slopes represent change over time. LBH589 Both are adjusted for covariates and fitted as random effects allowing each to vary between individuals.

The equation has three parts. Where t = time score (0, 1 or 2), i = individual,/γ = outcome, x = time score, η0 = intercept, η1 = slope, x/w = time invariant-covariate, α = factor loadings for the intercept, γ = factor Pazopanib loadings for the slope, and ε/ζ = residuals: (1) yti = η0i + η1ixt + εti; (2) η0i = α0 + γ0wi + ζ0i; (3) η1i = α1 + γ1wi + ζ1i. In the structural equation modelling framework, equation (1) is the measurement part, defining factor loadings that determine the shape of the growth factors and equations (2, 3) are the structural part, determining regressions among latent variables and on covariates ( Kline, 2011). The latent variable for the intercept represents initial status, the estimated value of the outcome at time score zero. The latent variable for the slope represents the expected linear increase

in the outcome as the time score changes from zero to one, given that time scores are coded 0, 1, 2 ( Bollen and Curran, 2006 and Duncan and Duncan, 2004). For the main analysis, we used multivariate (parallel process) LGC models (Bollen and Curran, 2006) to examine cross-sectional, longitudinal and bidirectional either associations between two growth processes simultaneously: mental health and physical activity. The regressions of the physical activity slope on the mental health intercept and the regression of the mental health slope on the physical activity intercept represent bidirectional effects (if the starting point of one predicts change in the other). The correlation between intercepts represents the estimated correlation at baseline. The correlation between slopes represents a bidirectional effect (both variables ‘moving together’ over time). The main advantage of this approach is that correlations between the starting point and change in two outcomes are modelled simultaneously. Several sensitivity analyses were conducted.